<h1 id="beanflk-masturbation-lora-with-audio">BEANFLK — Masturbation LoRA with Audio</h1>Built for one thing most video models won't do: her fingers actually going in.Trigger: <code>BEANFLK</code> — all caps, first word of your prompt.<h2 id="why-this-exists">Why this exists</h2>Ask almost any video model for a woman fingering herself and you'll get a hand rubbing around on the outside, forever. It'll look fine. She'll never actually go in.That's the problem this LoRA was built to solve. Fingers pushing in deep until her knuckles press against her, working inside her, and drawing back out wet. That's the whole point of it — everything else here is built around making that one thing land.The dataset was assembled specifically for it: entry and withdrawal balanced clip for clip, every window cut so the insertion finishes inside the frame instead of getting trimmed off halfway through, and a bucket of clean anatomy footage so the model knows what it's drawing when the hand moves away.<h2 id="what-else-it-does">What else it does</h2><ul><li>Audio — wet sounds, breathing, moans and a real climax, generated with the video rather than dubbed on.</li><li>Climax — hips lifting and bucking, her pussy clenching in waves. Trained as its own thing, not just a hand moving faster.</li><li>Depth control — fingertips barely dipping in, or all the way to the knuckle. Both are trained, and you can ask for either.</li><li>Rubbing and teasing — the build-up, so it isn't only good at the main event.</li><li>Expression — she reads as enjoying it, rather than operating a hand.</li></ul><h2 id="two-versions-and-they-are-not-equal">Two versions — and they are not equal</h2>MiniMax H3 — use this one.<ul><li>Insertion fires at 0.40. Normal strength, no overdriving, no artifacts.</li><li>It will start the action for you. Give it a frame where her hands are up by her chest and it brings them down and begins the whole thing on its own.</li><li>Rebuilt from scratch: 1,775 clips across 14 sources, entry and withdrawal balanced clip for clip, every window cut so the insertion finishes inside the frame.</li><li>Climax trained as its own thing — hips bucking and lifting, not just a hand moving faster.</li><li>Audio trained jointly with the video in the same pass.</li></ul>LTX-2.3 / Eros — the original.<ul><li>Insertion needs 1.75, and only works if the source image already has her hand at or above the opening. It won't start the action from a cold frame.</li><li>Push much past 1.69 and artifacts creep in — nipples moving faster than the breast, breast growth on flat-chested subjects.</li><li>The audio and the rubbing are genuinely good, and it's still the right choice if LTX is your pipeline.</li></ul>Straight answer: the H3 version is better at the thing this LoRA exists for, and it isn't close. The LTX build was made first, before a lot of what we learned about cutting the dataset for insertion. If you can run H3, run H3.An LTX update is coming. Everything that made the H3 version work — the rebalanced dataset, the recut windows, the dedicated climax and anatomy footage — applies to LTX too. It just hasn't been baked yet. If you're on LTX, this page is worth watching.Each version box carries its own settings. The numbers do not transfer. LTX runs strengths above 1.0; H3 runs well under it.<h2 id="prompting">Prompting</h2>You are not limited to what it was trained on. This sits on top of a capable base model — ask for a different setting, a different position, her licking her fingers afterwards, and the base handles it while the LoRA supplies the action and the sound.Put BEANFLK first, then describe the scene and the action. The H3 version box has a full word list showing which phrases pull which behaviour.<h2 id="a-note-on-close-ups">A note on close-ups</h2>At extreme macro framing the base model's own anatomy becomes the limiting factor rather than this LoRA. Each version box says what to stack for that platform. At normal and wide framing you don't need anything extra.Every gallery clip lists its full prompt and the exact LoRA stack that made it.

BEANFLK — Masturbation & Finger Insertion + Audio

The original release. Trained on LTX-2.3 / Eros with both video and audio branches, so it generates its own sound rather than leaning on the base model.Quick start<ul><li>Base model: Eros v1.4 (LTX-2.3 22B)</li><li>Text encoder: Gemma 3 12B IT</li><li>Loader: LTX2 LoRA Loader Advanced</li><li>Strength: 1.4–1.69 general, 1.75 for reliable insertion</li><li>Branch weights: video 1, audio 1, v2a 1, a2v 1, other 1</li><li>Resolution: 800x800, match your source image's aspect</li><li>Duration: 10s @ 25fps</li><li>Speed LoRA: ltx-2.3-22b-distilled-lora-384-1.1 at 0.5 — keep it on</li></ul>Strength — read this bit<ul><li>1.0 — too weak. Her hand hovers just off the skin. No real contact, and she doesn't read as enjoying it.</li><li>1.4 to 1.69 — the general working band. Clean rubbing with real contact and expression.</li><li>1.75 — reliable insertion, but only with a properly primed source image. See below.</li><li>Above ~1.69 — artifacts creep in: nipples wiggle faster than the breast itself, and on a flat-chested subject her breasts start to grow.</li></ul>It's a gradient, not a wall. Plenty of renders above 1.69 are fine — just know what you're looking for.Your source image matters as much as strengthFor insertion, her hand must be at least directly above the opening. Any further away and insertion is not guaranteed — and cranking the strength will not rescue it. You'll get surface rubbing at a higher strength plus the artifacts that come with it.Fix the frame, not the slider.Rubbing has no such requirement. It works from almost any starting pose.PromptingPut BEANFLK first, then describe the scene and action in plain prose. Keep the audio description in the same sentence flow rather than a separate block.Example: BEANFLK pale woman, tight crop on her pussy and hand, pushing her fingertips deep into her vagina, wet sounds of her pussy with moans and sighsKeep the word deep if you want depth. It's deliberately absent from the training captions, so the base model supplies it and the two don't fight.Sampler settings<ul><li>Pass 1: euler_cfg_pp, sigmas 0.85, 0.4219, 0.0</li><li>Pass 2: euler_ancestral_cfg_pp, sigmas 1.0, 0.9, 0.72, 0.4, 0.0</li><li>CFG 1.2 both passes</li><li>Upscaler: ltx-2.3-spatial-upscaler-x2-1.1</li></ul>Don't set the speed LoRA to 0. Those 2-step and 4-step schedules only work because it's engaged. Turn it off without raising your step count and you get blurry, vibrating motion.

Files with SHA256: 8e915c0336c78ed47f65fe710848ae4b0917d0e3579a29fab55c32f1e371d1a5